Efficient Multi-task Adaption for Crossbar-based In-Memory Computing

Fan Zhang, Li Yang, Deliang Fan

Abstract: ReRAM crossbar Non-volatile memory (NVM) based In-Memory Computing (IMC) has been widely investigated as a highly parallel, fast, and energy-efficient computing platform for Deep Neural Networks (DNNs), especially for one specific task inference. However, due to the intrinsic high energy consumption of weight re-programming and the relatively low endurance issue, adapting the ReRAM crossbar-based IMC hardware for continual learning or multi-task learning has not been well explored. In this paper, we discuss a crossbar-aware learning method with a 2-tier masking technique that could enable efficient and fast new task adaption for a deployed DNN model with minor hardware overhead.